Partially Identified Stan Model of COVID-19 Spread

Robert Kubinec writes:

I am working with a team collecting government responses to the coronavirus epidemic. As part of that, I’ve designed a Stan time-varying latent variable model of COVID-19 spread that only uses observed tests and cases. I show while it is impossible to know the true number of infected cases, we can rank/sign identify the effects of government policies on the virus spread. I do some preliminary analysis with the dates of emergency declarations of US states to show that states which declared earlier seem to have lower total infection rates (though they have not yet flattened the infection curve).

Furthermore, by incorporating informative priors from SEIR/SIR models, it is possible to identify the scale of the latent variable and provide more informative estimates of total infected. These estimates (conditional on a lower bound based on SIR/SEIR models) report that approximately 700,000 Americans have been infected as of yesterday, or roughly 6-7 times the observed case count, as many SEIR/SIR models have predicted.

I’m emailing you as I would love feedback on the model as well as to share it with others who may be engaged in similar modeling tasks.

Paper link

Github with Data & Stan code